How Learning Can Guide Evolution of Communication Reiji Suzuki and Takaya Arita Graduate School of Information Science, Nagoya University, Furo-cho, Chikusa-ku, Nagoya 464-8601, Japan {reiji, arita}@nagoya-u.jp Abstract The Baldwin effect is known as a possible scenario of the ge- netic acquisition process of a learned trait without the Lamar- ckian mechanism. However, it is still controversial how learn- ing can facilitate evolution in dynamically changing environ- ments caused by internal factors. Our purpose is to clarify whether and how leaning can facilitate evolution in dynamic environments which arise from communicative interactions among individuals. We constructed a simple computational model for the evolution of communication ability and its phe- notypic plasticity. In the model, the levels of adaptive com- munication, which correspond to the expected fitness value when the communication results in success, of signalling and receiving processes are determined by different sets of traits under the assumption of the correlation between their fitness and the effects of epistatic interactions among traits. A com- munication is successful only when the levels of the signaller and the receiver are the same, and the individuals try to im- prove their communication levels through the learning pro- cess in which the values of plastic traits can be modified from their genetically determined values. The evolutionary exper- iments clearly showed that the Baldwin effect repeatedly oc- curred and facilitated the adaptive evolution of communica- tion in this type of dynamic environments. Introduction The Baldwin effect (Baldwin, 1896, 1902) and the role of phenotypic plasticity in evolution have been drawing much attention in evolutionary studies (West-Eberhard, 2003; Crispo, 2007). The Baldwin effect is typically interpreted as a two-step evolution of the genetic acquisition of a learned trait without the Lamarckian mechanism: individuals that have successfully adapted their own trait to the environment through their lifetime learning processes occupy the popula- tion (1st step), and then the evolutionary path finds the innate trait that can replace the learned trait (2nd step) because of the cost of learning (Turney et al., 1996; Maynard-Smith, 1987). The second step is also known as genetic assimila- tion (Waddington, 1953), or a kind of genetic accommoda- tion (West-Eberhard, 2003; Crispo, 2007). Since the study by Hinton and Nowlan (Hinton and Nowlan, 1987), the computational approaches on this effect have contributed to understanding of how learning can af- fect evolution. An important finding of these studies is that the balances between the benefit and cost of learning can smooth the fitness landscape, and as a result, can either facil- itate or slow down the adaptive evolution. Especially, it has been reported that there can be situations in which learn- ing is not always beneficial for genetic evolution (Mayley, 1997; Paenke et al., 2006). For example, if there is no cost for learning an adaptive trait, there is no difference in the fitness between the learned one and the genetically acquired one. In this case, the learning behavior can retard the genetic evolution of such a trait because the selection pressure can- not distinguish between these traits. Thus, it is an important issue how learning can become necessary or unnecessary for adaptive genetic evolution depending on various states of a population and its environment. Recently, we discussed whether and how learning can fa- cilitate the adaptive evolution of population on rugged fit- ness landscapes (Suzuki and Arita, 2007b). We constructed a simple fitness function that represents a multi-modal fit- ness landscape as typically illustrated in Fig. 1, in which there is a correlation between the adaptivity of individual and the effects of epistatic interactions among its traits. The evolutionary experiments of the individual traits and their phenotypic plasticity on this landscape clearly showed that the Baldwin effect repeatedly occurred through the evolu- tionary process of the population, and facilitated its adaptive evolution as a whole. Also, the effects of learning on evolution have been dis- cussed in the context of dynamically changing fitness land- scapes. In such situations, we can expect that more complex scenarios of interactions between evolution and learning emerge because the balances between the benefit and cost of learning also change dynamically. While several studies focused on the effects of changes in the environmental con- ditions caused by the external factors (Sasaki and Tokoro, 1997; Ancel, 1999), we can also assume more complex sit- uations in which the fitness landscapes can be changed by internal factors (Suzuki and Arita, 2004). The evolution and emergence of communication is one of the typical cases of Artificial Life XI 2008 608